TL;DR
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
Software engineer Jesse Waites reports using a multi-stage AI process to search millions of historical records and identify candidate evidence of a meteorite fall, three rhinoceroses and volcanic eruptions. He says he checked the candidates against scans of original documents and specialist catalogues; the findings are presented as research leads, not as a comprehensive independent verification of every historical claim.
Software engineer Jesse Waites says he used AI to search millions of digitized historical records, finding a report of a forgotten meteorite, references to three rhinoceroses and evidence that may help identify previously unrecorded volcanic eruptions. The work matters because it shows how automated screening can direct researchers to records buried in vast archives, though each candidate still requires checking against original documents and specialist research.
Waites describes building a research pipeline after reading historian Benjamin Breen’s account of an AI-assisted search that identified a 1615 ship’s journal mentioning dodos on Mauritius. Breen’s work used digitized records of the Dutch East India Company; Waites adapted the general approach to investigate several historical questions in parallel.
The sources Waites says he searched included 4.35 million pages of Dutch East India Company material transcribed by the GLOBALISE project, digitized Dutch newspapers, two centuries of American newspapers and selected ship logbooks. He reports that the system converted 5.7 million passages from the Company archive into meaning-based representations, helping it find relevant material despite inconsistent historical spelling and errors in text recognition.
His process used a fast, lower-cost AI model to screen candidate passages, then a more capable model to assess a smaller set, translate material and extract dates and places. Waites says he checked promising results against scans of the original handwritten or printed pages and consulted relevant specialist catalogues before describing them as new. His report does not provide independent confirmation that every candidate changes the accepted historical record.
AI Can Narrow the Archive Search
The project illustrates a practical role for AI in historical research: finding and prioritizing records, rather than replacing the work of historians or archivists. Large collections can contain relevant evidence that is difficult to retrieve through ordinary keyword searches, especially when spelling varies or optical character recognition has misread the page.
Waites says he used an inexpensive screening step before asking a larger model to examine likely matches. He gives the example of screening 59,000 elephant mentions for about $3 in model costs. That is a figure reported by Waites for this workflow, not a general cost estimate for archive research or a measure of the quality of the results.
The discoveries could offer researchers new leads on animal histories, meteorite records and volcanic events. Their wider value depends on whether specialists can verify the readings, establish the documents’ provenance and determine whether the material has already been noted in other catalogues or scholarship.
As an affiliate, we earn on qualifying purchases.
From Dodo Journal to Wider Searches
The immediate inspiration was Breen’s report, published on October 1, 2026, about a 1615 journal in the Dutch East India Company archive. As described in the supplied material, the journal records sailors on Mauritius catching tortoises and dodos. The Company operated from 1602 until its dissolution in 1799, leaving a large body of records that is now being digitized and transcribed.
Waites sought questions that could be investigated using material already available online and whose answers could be checked against an original page. His candidate topics included animals in Company records, a large volcanic eruption in 1808 that has not been located, earthquakes reported in old Dutch newspapers, meteorite falls and missing ships.
Historical handwriting and older print create a particular search problem: the same word may appear in many spellings, and machine-generated transcriptions can contain errors. Waites’s use of meaning-based search was intended to find passages based on their subject, not only on an exact word match. The AI results were then treated as pointers back to archival evidence.
“Adding even one new data point to a couple of niche fields felt like a small but worthwhile contribution to make.”
— Jesse Waites, describing his research goals
Which Findings Will Hold Up
The supplied account does not include enough detail to independently evaluate the meteorite report, the three rhinoceros records or the proposed volcanic evidence. It does not identify all the documents, dates, archive references or catalogue checks needed for readers to retrace each finding from this material alone.
It is also unclear whether every result is entirely new to specialists, whether the records establish the events as described, or whether any readings remain disputed. A document can provide evidence of an observation without resolving its interpretation, and AI-generated translations or summaries can introduce errors. Waites says he examined original-page scans, but independent review of the individual records is not described here.
The report’s figures for archive size, passages searched and model costs are Waites’s descriptions of his own project. They do not by themselves show how complete the search was or how often the system missed relevant material.
Follow the Records to Review
The next step is for researchers and readers to examine the cited scans and archive references, then compare the candidate records with existing catalogues and scholarship. That review can establish whether the material represents a genuinely overlooked observation, supports a revised chronology or has already been documented elsewhere.
Waites’s account describes a method that could be applied to other digitized collections, but the results will depend on transcription quality, search design and human verification. The supplied material does not announce a formal peer-reviewed publication or set out a schedule for independent assessment of the findings.
Key Questions
What did the AI-assisted search find?
Jesse Waites reports finding a meteorite report, references to three rhinoceroses and records that may help identify volcanic eruptions. The supplied account does not establish that all the findings have been independently confirmed.
Which archives did Waites search?
He says he searched GLOBALISE transcriptions of Dutch East India Company records, digitized Dutch newspapers, two centuries of American newspapers and selected ship logbooks.
Did AI verify the historical discoveries?
No. Waites describes AI as a way to locate and screen possible records. He says he checked promising passages against scans of original pages and specialist catalogues; the historical significance still depends on verification and interpretation.
How did the search handle old spelling and transcription errors?
Waites says the system used meaning-based search to find passages by subject rather than relying only on exact keywords, then used AI models to screen and interpret candidates. He reports checking selected results against original-page scans.
What remains unknown about the findings?
The supplied report does not give enough document-level information to independently assess every result, and it is unclear whether all the records are new to specialists or how they will affect existing historical accounts.
Source: hn
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
